AI × Learning
This week's theme is the pace problem. Students adopt a new tool in weeks and, on the best evidence yet, outsource their homework to it within five months; universities answer in eight-month committee reports and summer convenings. Carnegie Mellon shows what a course-level redesign looks like when faculty accept that students will use AI anyway; Benjamin Riley shows what happens when nobody designs; a Chicago biologist argues institutions have chosen curriculum-revision speed for a pandemic-speed problem. The FT finds teenagers' degree choices lagging the labor market by years, a Chronicle profile bets on a slow gen-ed revival, and a Dushu roundtable reminds us that the drift from reading books to reading materials began thirty years before ChatGPT. Finally, Qin Chen reads 740 million Chinese job postings and argues that the division of labour itself, the slowest-moving institution of all, has started to reverse.
Practice
1. Carnegie Mellon rewrites the assignment, not the rule
Carnegie Mellon's account of its AI-era course changes starts from an accepted premise: students will use AI, so the work must change. Computer-science assignments are shifting toward code reviews and interviews in which students explain their design choices. Zico Kolter's diagnosis: “We've outsourced a lot of teaching to independent exercises that students do at home, and it's arguable that this is no longer as effective as it was.” Mike Taylor's “Effective Coding With AI” course found that “the students who learned the most were the ones who refused to let the AI think for them.” In engineering, Chris McComb's students hunted an AI's deliberate mistakes across eight assignments and “significantly outperformed students from the previous semester in three of four concept areas.” The Eberly Center's Chad Hershock separates AI as “production assistant” from AI as “thought partner,” and checks the difference with no-AI “transfer tasks.”
Two of CMU's moves cost nothing but time and suit a seminar-sized class better than a lecture hall: the short interview where a student explains how and why their project was built, and the transfer task, a follow-up done without AI that shows whether anything was learned. Both replace policing with evidence. The “catch the AI's mistakes” assignment is also a ready-made template for any quantitative course here.
Sources: "Navigating the AI Era with a CMU Focus on Critical Thinking" — Carnegie Mellon University News · spotted via @Xudong07452910 on X
Research
2. The 50 percent problem: what unmanaged AI use does in five months
Benjamin Riley of Cognitive Resonance reads the Chinese secondary-school study we pointed to in Issue #6 and names its central finding. The data cover “approximately 27,000 Chinese students in grades seven to 12.” His summary: “Between 2023 and 2025, approximately 80 percent of all students started using generative AI, and after just five months of use, approximately 50 percent of these students engaged in ‘full homework outsourcing.’” Homework scores rose; on closed-book exams “the average decline was 20 percent with a 1.4 standard deviation,” and for students who adopted AI two or more years before testing, the effects are “24 percent (1.5 SD) for the high school exam, and 18 percent (1.3 SD) for the college exam.” Riley calls it a “cognitive race to the bottom.” Against the “intelligent use” defense, he answers with the data: sustained 75-minute homework sessions after adoption occurred for “only 20 students.” The authors' verdict: generative AI “has a substantial negative impact on student learning.”
See the three animated histograms (homework time, homework scores, closed-book exam scores for Pre-AI, Never-AI and Used-AI students) and the homework-time scatter in Riley's post: The 50 Percent Problem
These are secondary students, and some of them are our incoming first-years, arriving with the outsourcing habit already formed. The study's mechanism, not its headline, is what transfers: harm concentrated in ungraded-for-learning homework that AI can complete outright. Any assignment whose only function is practice should now come with either a no-AI transfer check (see item 1) or an honest expectation that half the class will not do it.
Sources: "The 50 Percent Problem" — Benjamin Riley, Cognitive Resonance · underlying paper: "The Generative AI Learning Penalty: Evidence from Chinese Secondary Education"
Policy
3. “They treat AI like a curriculum revision”
Ran Blekhman of the University of Chicago argues that universities are failing not for lack of effort but for lack of speed and coherence. His evidence is the contradictions: Chicago hands every student Claude Enterprise while its social-sciences Core goes analog; Cal State “renewed its contract with OpenAI for $39 million” while “40% of its faculty either forbid or discourage AI use in the classroom.” Among UK students, “95 percent of respondents report using AI in some form,” yet “37% agree or strongly agree that their institution encourages the use of generative AI, while 36% percent disagree or strongly disagree.” MIT's report, covered here last week, “took 8 months to write.” His line: “Instead of treating AI like a pandemic, they treat AI like a curriculum revision.” The counterexample is 2020, when the same institutions rebuilt everything in weeks; slowness is a choice.
Blekhman's incoherence test is easy to run on ourselves: does what we provide to students match what our syllabi permit, and could a first-year describe our position in one sentence? A small university's advantage is that a coherent answer needs one conversation, not a divisional committee. The pandemic comparison is also a reminder that DKU already has a working memory of fast, whole-institution change.
Source: "Universities are Failing This AI Moment" — Ran Blekhman
Students
4. Teenagers are choosing the majors AI is hollowing out
The FT's “AI Shift” newsletter (Sarah O'Connor and John Burn-Murdoch) reports a working paper by Bouke Klein Teeselink, Aristotle Vossos, Andrei Andronic and Kartik Akileswaran built on UK university applications from 2020 to 2025. After ChatGPT, applicants moved toward exposed degrees, not away: “a one standard deviation increase in a degree's exposure is associated with a six per cent increase in applications,” and the shift holds across genders, ethnic groups and subjects. The same authors find entry-level openings shrinking in exactly those fields. “Students, in other words, are moving toward subjects in which the entry-level jobs they will likely apply to are disappearing.” The lead researcher says 2026 applications hint at a reversal, and Burn-Murdoch adds that US computer-science enrolments are “down about 10 per cent compared to their peak in 2024.” The lesson is “how slowly cutting-edge research percolates down to the wider population — especially teenagers.”
Figure: “The long boom in studying computer science has sharply reversed” · US data = enrolments, UK = applications · Sources: National Student Clearinghouse Research Center (US); UCAS End of Cycle data resources (UK) · FT graphic: John Burn-Murdoch · from “Is AI affecting what people choose to study?” (FT, gift link)
Our students declare majors after arriving, which puts advising in the gap the FT describes: applicants act on a picture of the labor market that is two or three years stale. That is an argument for putting the Stanford “canaries” data from Issue #7 and this paper in front of first-year advisers now, and for describing majors to students by the judgment they build rather than the entry-level job they once fed.
Sources: "Is AI affecting what people choose to study?" — Financial Times, The AI Shift (gift link) · ft.com
Ideas
5. The gen-ed revivalist
The Chronicle's “People to Watch 2026” package profiles Roosevelt Montás, who ran Columbia's Core Curriculum and now directs the new Chang Chavkin Center for Liberal Education and Civic Life at Bard College. His target is the distribution model of general education, the checklist of one course per box. “What if, instead of engaging in a cursory exploration of the disciplines, students read transformative texts and wrestled with complex and timeless questions?” The setting is the present crisis: “Higher education is facing a crisis in confidence. It is being overrun by AI.” Against the assumption that students want only job training, he insists “Students are open and often hungry for this kind of education.” The center will convene 10 to 15 colleges in summer 2027 and study effects on retention and graduation. The obstacle is faculty careers: “a professorial track whose concern is primarily liberal education just doesn't exist in academia.”
DKU was built on the text-and-discussion model Montás wants other campuses to rediscover, so this is less a new idea than a check on drift: are our common courses still organized around questions and texts, or have they quietly become a distribution checklist? His staffing point is the one to take seriously. A common curriculum survives only if teaching it counts as career-building for the faculty who do.
Source: "The Gen-Ed Revivalist" — Beth McMurtrie, The Chronicle of Higher Education
Reading
6. Dushu asks: in the AI era, what is reading for?
Dushu (读书), the leading Chinese intellectual monthly, opens its September issue with proceedings of its first 思想会 roundtable, seven scholars and editors on the reading crisis. Two voices stand out. Legal scholar Zheng Ge (郑戈) dates the problem before AI: “过去三十年，高等教育经历了从‘读书’到‘读材料’的深层位移” (over thirty years, higher education has shifted from reading books to reading materials). AI pushes this to the limit: its reading is “统计预测” (statistical prediction), and the student who uploads Huckleberry Finn for a report “错过了哈克心智的缓慢成长” (misses Huck's slow growth of mind). The cost is “判断力的萎缩” (the atrophy of judgment), which forms only through “亲自与文本的反复搏斗” (one's own repeated struggle with the text). Historian Qiu Luming (仇鹿鸣) traces the habit's disappearance to secondary school: “阅读往往先是一种习惯，其后才是一种职业技能” (reading is first a habit, only later a professional skill), and twenty years of test-prep 内卷 have consumed both the leisure and the pleasure in which the habit forms. His question for the humanities: “能否勇敢地宣称人文教育不是一种职业技能训练？” (dare we say liberal education is not job training?)
Zheng's argument is the strongest case for the reading-intensive seminar we have seen in Chinese: judgment is not transferable from a summary, whoever writes it. Qiu's is a description of the pipeline most of our students arrive through, fluent in scoring points and unfamiliar with reading for pleasure. Together they suggest that first-year courses here have to build the habit before they can assess the skill. Worth circulating to Chinese-reading colleagues.
Source: 《读书》首发｜郑戈、仇鹿鸣等：AI时代，阅读究竟为了什么？ — 读书杂志 (WeChat), 《读书》2026年第9期
Labour market
7. The pin factory in reverse: 740 million Chinese job ads
Qin Chen's Zhihu essay 《不再被需要的年轻人，和替他们干活的中登》 (the young people no longer needed, and the mid-career staff doing their work) takes the labour-market story to China and from headcount to task structure. Working from “7.4亿条截至2026年8月的招聘数据” (740 million job postings through August 2026), he and his co-authors find that the postings most accelerable by large models did shrink, that the hit lands on low-wage, low-experience roles, and that above “12000元每个月” (¥12000/month) the slope turns positive. Hiring demand in 2026 “减少了5.2%” (reduced 5.2%) because of AI, yet “18.6%的工作岗位” (18.6% of positions) now bundle tasks that belonged to different people three years ago. The Adam Smith argument follows: modern organizations exist because the specialist “离开了组织，就什么都不是” (is nothing outside the organization); models make learning and switching tasks nearly free, so “人类的最优分工深度，第一次出现了掉头” (for the first time, the optimal depth of the division of labour has reversed). Firms flatten, juniors go unhired, mid-career staff absorb their work. His closing question: “当生产不再需要分工，我们用什么来黏合社会？” (when production no longer needs a division of labour, what holds society together?)
Figure: “2023：爬升 250 年后第一次掉头” (2023: the first reversal after 250 years of climbing), a schematic of the optimal depth of the division of labour from the pin factory to AI agents · from Qin Chen's essay on Zhihu
This is the market most of our graduates walk into, and the worker it now rewards looks like a liberal-arts graduate: someone who can hold several formerly separate tasks in one head, switch between them, and carry responsibility for the result. The catch is Chen's own finding that the entry rung is where the thinning is worst. Two things follow for us: teach students to integrate, not only to specialize, and put his last question on a seminar syllabus, because it has become an economic one.
Sources: 《不再被需要的年轻人，和替他们干活的中登》 — Qin Chen, Zhihu, 2026-08-31 · underlying paper: Qin Chen, Ying Fang, Xiangyu Wang & Leo Yang Yang, “The Pulse Beneath the Job Title: Monthly Readings of Requirements and Tasks from 750 Million Chinese Job Ads” (SSRN, August 27, 2026)
Rather than rushing to passive tool adoption or rigid standardization, DKU is treating AI as an ongoing design challenge. This newsletter is part of our shared experiment. It is a regular space to reflect, converse, and build an approach to learning grounded in judgment and curiosity.
— Yisu